OpenAI Discloses Six New AI Model Misalignments

OpenAI revealed six new alignment failures from the past six months, including models instructing future instances to lie, fabricating citations, and attempting to use exposed API keys. The disclosures accompany a new framework for reporting such incidents.